A knowledge and data fusion-based power system residual life prediction method
By employing a knowledge and data fusion approach, utilizing Wiener processes and convolutional gated recurrent neural networks, and combining Monte Carlo and stacking strategies, the accuracy and interpretability issues of power system lifetime prediction were addressed, resulting in more accurate remaining lifetime prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2022-10-25
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies struggle to accurately predict the lifespan of power systems, especially in complex equipment. Methods based on physical and statistical models lack domain knowledge and are inaccurate, while data-driven methods struggle to quantify individual degradation conditions and have poor interpretability.
By employing a knowledge and data fusion approach, historical data of the power system is accumulated to establish a statistical model based on the Wiener process and a convolutional gated recurrent neural network model. Combined with Monte Carlo methods and stacking fusion strategies, a neural network model is constructed to predict the remaining lifetime.
It improves the accuracy and interpretability of power system lifetime prediction by generating a more accurate remaining lifetime distribution by fusing results from multiple models.
Smart Images
Figure CN115577637B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system lifetime prediction technology, and more particularly to a method for predicting the remaining lifetime of a power system by integrating knowledge and data. Background Technology
[0002] Power systems involve numerous monitoring parameters. Composed of individual battery cells, their spatial dimensions range from micrometers to meters. Over time, they undergo cyclic charging and discharging processes. Furthermore, the electrical and thermal stresses experienced by power systems vary, making lifespan prediction challenging. As power systems are complex equipment, accurate lifespan prediction provides crucial information for health management, ensuring the effective operation of such equipment.
[0003] Power system degradation models based on physical and statistical models are difficult to accurately describe the degradation process of power systems, and suffer from insufficient domain knowledge and complex model building. Although data-driven power system degradation models utilize a large amount of historical data, break through the knowledge bottleneck, and achieve more accurate predictions, they are difficult to quantify the degradation status of individual systems and have poor interpretability.
[0004] Therefore, how to provide a knowledge and data fusion method to improve the accuracy of power system lifetime prediction has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] The purpose of this invention is to provide a method for predicting the remaining lifetime of a power system by fusing knowledge and data, in order to solve the above-mentioned problems.
[0006] The present invention solves the technical problem by adopting the following technical solution:
[0007] A method for predicting the remaining lifetime of a power system by fusing knowledge and data includes the following steps:
[0008] Step S1: Accumulate timing data from the historical operation of the power system and from the cycle life test;
[0009] Step S2: Preprocess and extract features from the time-series data of the power system during its historical operation to obtain data with degradation trends;
[0010] Step S3: Establish a power system capacity degradation model based on the Wiener process, which is the statistical model;
[0011] Step S4: Based on the accumulated data on the power system capacity degradation trend, solve for the parameters in the power system capacity degradation model;
[0012] Step S5: Calculate the first arrival time and the remaining lifetime distribution of the power system based on the industry power system capacity failure threshold.
[0013] Step S6: Construct a degradation model based on a convolutional gated recurrent neural network, with historical monitoring data as input and remaining lifetime as output;
[0014] Step S7: Train a degradation model based on a convolutional gated recurrent neural network to obtain a data-driven model, and estimate the remaining lifetime distribution using the Monte Carlo method;
[0015] Step S8: Based on the stacking fusion strategy, construct a neural network model that fuses knowledge and data;
[0016] Step S9: Integrate the relevant results of the statistical model obtained in steps S3-S5 and the data-driven model obtained in steps S6-S7, and use the integrated sample data as input to train the neural network model of knowledge and data fusion.
[0017] Step S10: For the newly input sample data, update the statistical model online and estimate the remaining lifetime distribution, and use the trained data-driven model to predict the remaining lifetime distribution; use the trained knowledge and data fusion neural network model to fuse the results of the two.
[0018] Step S11: Based on the fusion results in step S10, calculate the weights and sample the statistical model and the data-driven model respectively to obtain a more accurate remaining lifetime distribution.
[0019] Furthermore, step S2 specifically includes:
[0020] Preprocessing of time series data includes filling in missing time series data and normalization.
[0021] The variance statistic is used to determine whether a performance parameter has a trend of change. The variance calculation formula is: Var(X) = E[XE(X)] 2 X is the variable to be judged. Based on the charging and discharging principle of the power system, the changes in the parameter curves are observed, and the performance parameters with degradation trends are selected as the modeling objects.
[0022] Furthermore, step S3 specifically includes:
[0023] Construct a power system capacity degradation model based on the Wiener process: Y(t)=y0+λt+σ B ·B(t);
[0024] The power system capacity degradation model is discretized and transformed into the following state equation form:
[0025]
[0026] In the formula, λ is the drift coefficient of the degradation process, λ0 follows a normal distribution N(μ0,σ0), where μ0 and σ0 are latent variables to be estimated; ε is the noise of the random walk process, following a normal distribution N(0,Q), where Q is a latent variable to be estimated; y is the degradation amount under study; t is the time of monitoring data; σ B is the coefficient of Brownian motion, is the latent variable to be estimated; B is Brownian motion.
[0027] Furthermore, step S4 specifically includes:
[0028] The drift coefficients are estimated using the Kalman filter algorithm;
[0029] The parameters in the power system capacity degradation model are solved iteratively using the EM algorithm, where the maximum likelihood formula is:
[0030]
[0031] In the formula, θ=[μ0,σ0,Q,σ B [Y] represents the latent variable to be estimated; 0:i =[y0,y1,…y i ] represents the set of degradation values y at each time step; Λ i =[λ0,λ1,…,λ i ] represents the drift coefficient λ at each time point. i The set of values; finally, the power system capacity degradation model based on the Wiener process is solved.
[0032] Furthermore, step S5 specifically includes:
[0033] The power system failure threshold is 30% capacity loss. Based on the definition of first-arrival time and the inverse Gaussian distribution, the lifetime distribution is calculated as follows: In the formula R i Let r be the set of remaining lifetimes. i Let w be a certain estimated value of the remaining lifetime, and w be the failure threshold of the power system capacity. Substitute the relevant variables from the power system capacity degradation model into the equation to obtain the remaining lifetime distribution.
[0034] Furthermore, step S6 specifically includes:
[0035] The network input consists of voltage, current, and temperature data from historical monitoring data, and the network output is the remaining lifetime.
[0036] The input time step is m, and the remaining lifetime is predicted using convolutional layers and gated recurrent neural network layers;
[0037] The hidden layers of the network consist of convolutional neural networks, gated recurrent neural networks, and fully connected layers.
[0038] Furthermore, step S7 specifically includes:
[0039] Based on the degradation model derived from S6 using a convolutional gated recurrent neural network, the lifetime distribution of the data-driven model is obtained through Monte Carlo dropout simulation.
[0040] Furthermore, step S8 specifically includes:
[0041] A neural network model for knowledge and data fusion is constructed based on convolutional neural networks and fully connected neural networks;
[0042] The network's inputs are the outputs of the statistical model and the data-driven model, and the network's output is the predicted mean of the remaining lifetime distribution.
[0043] The hidden layers of the network include convolutional neural networks and fully connected neural networks.
[0044] Furthermore, step 11 specifically includes:
[0045] Based on the neural network model that integrates knowledge and data, and combined with the complementary filter model, the sampling weights are calculated.
[0046] By sampling the statistical model and the data-driven model according to the weights, a more accurate distribution of remaining lifetime is obtained.
[0047] Beneficial effects:
[0048] This invention discloses a knowledge and data fusion method for predicting the remaining lifetime of a power system. It designs a capacity degradation model based on the Wiener process based on domain knowledge to estimate the remaining lifetime distribution of the power system; uses a data-driven approach to predict the remaining lifetime distribution of the power system based on a convolutional gated recurrent neural network; and designs a knowledge and data fusion model based on a stacking fusion strategy to obtain a more accurate remaining lifetime distribution. Attached Figure Description
[0049] Figure 1 A flowchart of the method for predicting the remaining lifetime of a power system provided by the present invention.
[0050] Figure 2 This is a structural diagram of the degenerate model based on a gated convolutional neural network in this invention.
[0051] Figure 3 This is a structural diagram of the knowledge and data fusion model designed based on the stacking fusion strategy of this invention.
[0052] Figure 4 This is a schematic diagram illustrating the dynamic weight generation of the remaining lifetime prediction distribution according to the present invention. Detailed Implementation
[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0054] Example 1
[0055] This invention discloses a method for predicting the remaining lifetime of a power system by fusing knowledge and data, comprising the following steps:
[0056] Step S1: Accumulate timing data from the historical operation of the power system and from the cycle life test;
[0057] Step S2: Preprocess and extract features from the time-series data of the power system during its historical operation to obtain data with degradation trends;
[0058] Step S3: Establish a power system capacity degradation model based on the Wiener process, which is the statistical model;
[0059] Step S4: Based on the accumulated data on the power system capacity degradation trend, solve for the parameters in the power system capacity degradation model;
[0060] Step S5: Calculate the first arrival time and the remaining lifetime distribution of the power system based on the industry power system capacity failure threshold.
[0061] Step S6: Construct a degradation model based on a convolutional gated recurrent neural network, with historical monitoring data as input and remaining lifetime as output;
[0062] Step S7: Train a degradation model based on a convolutional gated recurrent neural network to obtain a data-driven model, and estimate the remaining lifetime distribution using the Monte Carlo method;
[0063] Step S8: Based on the stacking fusion strategy, construct a neural network model that fuses knowledge and data;
[0064] Step S9: Integrate the relevant results of the statistical model obtained in steps S3-S5 and the data-driven model obtained in steps S6-S7, and use the integrated sample data as input to train the neural network model of knowledge and data fusion.
[0065] Step S10: For the newly input sample data, update the statistical model online and estimate the remaining lifetime distribution, and use the trained data-driven model to predict the remaining lifetime distribution; use the trained knowledge and data fusion neural network model to fuse the results of the two.
[0066] Step S11: Based on the fusion results in step S10, calculate the weights and sample the statistical model and the data-driven model respectively to obtain a more accurate remaining lifetime distribution.
[0067] In this embodiment, step S2 specifically includes:
[0068] Preprocessing of time series data includes filling in missing time series data and normalization.
[0069] The variance statistic is used to determine whether a performance parameter has a trend of change. The variance calculation formula is: Var(X) = E[XE(X)] 2 X is the variable to be judged. Based on the charging and discharging principle of the power system, the changes in the parameter curves are observed, and the performance parameters with degradation trends are selected as the modeling objects.
[0070] In this embodiment, step S3 specifically includes:
[0071] Construct a power system capacity degradation model based on the Wiener process: Y(t)=y0+λt+σ B ·B(t);
[0072] The power system capacity degradation model is discretized and transformed into the following state equation form:
[0073]
[0074] In the formula, λ is the drift coefficient of the degradation process, λ0 follows a normal distribution N(μ0,σ0), where μ0 and σ0 are latent variables to be estimated; ε is the noise of the random walk process, following a normal distribution N(0,Q), where Q is a latent variable to be estimated; y is the degradation amount under study; t is the time of monitoring data; σ B is the coefficient of Brownian motion, is the latent variable to be estimated; B is Brownian motion.
[0075] In this embodiment, step S4 specifically includes:
[0076] The drift coefficients are estimated using the Kalman filter algorithm;
[0077] The parameters in the power system capacity degradation model are solved iteratively using the EM algorithm, where the maximum likelihood formula is:
[0078]
[0079] In the formula, θ=[μ0,σ0,Q,σ B [Y] represents the latent variable to be estimated; 0:i =[y0,y1,…y i ] represents the set of degradation values y at each time step; Λ i =[λ0,λ1,…,λ i ] represents the drift coefficient λ at each time point. i The set of values; finally, the power system capacity degradation model based on the Wiener process is solved.
[0080] In this embodiment, step S5 specifically includes the following steps:
[0081] The power system failure threshold is 30% capacity loss. Based on the definition of first-arrival time and the inverse Gaussian distribution, the lifetime distribution is calculated as follows: In the formula R i Let r be the set of remaining lifetimes. i Let w be a certain estimated value of the remaining lifetime, and w be the failure threshold of the power system capacity. Substitute the relevant variables from the power system capacity degradation model into the equation to obtain the remaining lifetime distribution.
[0082] In this embodiment, step S6 specifically includes:
[0083] The network input consists of voltage, current, and temperature data from historical monitoring data, and the network output is the remaining lifetime.
[0084] The input time step is m, and the remaining lifetime is predicted using convolutional layers and gated recurrent neural network layers;
[0085] The hidden layers of the network consist of convolutional neural networks, gated recurrent neural networks, and fully connected layers.
[0086] In this embodiment, step S7 specifically includes:
[0087] Based on the degradation model derived from S6 using a convolutional gated recurrent neural network, the lifetime distribution of the data-driven model is obtained through Monte Carlo dropout simulation.
[0088] In this embodiment, step S8 specifically includes:
[0089] A neural network model for knowledge and data fusion is constructed based on convolutional neural networks and fully connected neural networks;
[0090] The network's inputs are the outputs of the statistical model and the data-driven model, and the network's output is the predicted mean of the remaining lifetime distribution.
[0091] The hidden layers of the network include convolutional neural networks and fully connected neural networks.
[0092] In this embodiment, step 11 specifically includes:
[0093] Based on the neural network model that integrates knowledge and data, and combined with the complementary filter model, the sampling weights are calculated.
[0094] By sampling the statistical model and the data-driven model according to the weights, a more accurate distribution of remaining lifetime is obtained.
[0095] This invention discloses a knowledge and data fusion method for predicting the remaining lifetime of a power system. It designs a capacity degradation model based on the Wiener process based on domain knowledge to estimate the remaining lifetime distribution of the power system; uses a data-driven approach to predict the remaining lifetime distribution of the power system based on a convolutional gated recurrent neural network; and designs a knowledge and data fusion model based on a stacking fusion strategy to obtain a more accurate remaining lifetime distribution.
[0096] Example 2
[0097] like Figure 1 As shown, this invention provides a method for predicting the remaining lifetime of a power system by fusing knowledge and data, comprising the following steps:
[0098] Step 1: Accumulate historical operational status data of the power system and collect real-time monitoring data of the power system. The monitoring data of the power system mainly includes information characterizing the equipment's operating status, such as charging current, charging voltage, charging temperature, discharging current, discharging voltage, and discharging temperature.
[0099] Step 2: Preprocess the historical data of the power system, including filling in missing time series data and normalization. Through feature extraction, obtain data with degradation trend, and calculate the capacity of each circulating power system through numerical integration.
[0100] Specifically, each step is explained in detail below:
[0101] Step 2.1: Preprocess the data, including missing data imputation and normalization.
[0102] (1) Missing data imputation: Fill with the average of the two nearest adjacent elements of the missing data;
[0103] (2) Data normalization processing: Where A is a time series, A i Let A be the i-th data in A. min Let A be the minimum value. max It is the maximum value in A.
[0104] Step 2.2: Determine whether the performance parameter has a trend based on the variance statistic, where the variance calculation formula is: Var(X)=E[XE(X)] 2 X is the variable to be judged. Based on the charging and discharging principle of the power system, the changes in the parameter curves are observed, and the performance parameters with degradation trends are selected as the modeling objects.
[0105] Step 2.3: The discharge capacity for each cycle is calculated using numerical integration on the data collected during the discharge phase. The specific calculation formula is as follows:
[0106]
[0107] In the formula Q c For power system capacity, I i Let t be the magnitude of the discharge current at time i. i This corresponds to the discharge duration.
[0108] Step 3: Establish a power system capacity degradation model based on Wiener processes.
[0109] Specifically, the detailed statistical degradation model is as follows:
[0110] Step 3.1: Use the capacity obtained from numerical integration in Step 2 as the input to the statistical degradation model;
[0111] Step 3.2: Construct a degradation model based on the Wiener process: Y(t)=y0+λt+σ B ·B(t)
[0112] Step 3.3: Discretize the model and transform it into the following state equation form:
[0113]
[0114] Where λ is the drift coefficient of the degradation process, σ B The coefficient for Brownian motion;
[0115] Step 4: Based on the accumulated power system capacity degradation data, solve for the parameters in the degradation model;
[0116] Specifically, the solution process for the degradation model is as follows:
[0117] Step 4.1: Estimate the drift coefficients using the Kalman filter algorithm; this includes four parts: initialization of latent variable parameters, recursive state update, error estimation, and update of the parameters to be estimated.
[0118] Step 4.2: Iteratively solve for each parameter in the Wiener process model using the EM algorithm, where the maximum likelihood formula is:
[0119]
[0120] Step 4.3, iterate through steps 4.1 and 4.2 until the estimated difference converges. Finally, the capacity degradation model based on the Wiener process can be solved. Substituting each latent variable into the model yields the corresponding model.
[0121] Step 5: Calculate the remaining lifetime distribution of the power system by solving the first arrival time based on the industry power system capacity failure threshold.
[0122] Specifically, the calculation process for lifetime distribution is as follows:
[0123] Step 5.1: The power system failure threshold is generally 30% capacity loss. Based on the definition of first-arrival time and the inverse Gaussian distribution, the lifetime distribution is calculated as follows:
[0124] Step 5.2: Substituting the latent variables obtained from the model solution into the equation yields the lifetime distribution expression.
[0125] Step 6: Construct a degradation model based on a convolutional gated recurrent neural network, with historical monitoring data as input and remaining lifetime as output.
[0126] Specifically, such as Figure 2 As shown, the established degradation model is explained in detail below:
[0127] Step 6.1: The network input consists of m-dimensional parameters, including selected monitoring data, whose time series can fully reflect the health status of the power system.
[0128] Step 6.2: Input a time step of n and use a time series of length n to predict the remaining lifetime of the power system.
[0129] Step 6.3: The hidden layers of the network are alternately connected by GRU layers, CNN layers and dropout layers. GRU layers can effectively extract temporal information from the data, CNN layers can capture short-range spatial relationships, and dropout layers are used to prevent overfitting.
[0130] Step 7: Train a degradation model based on a convolutional gated recurrent neural network and estimate the remaining lifetime distribution using the Monte Carlo method.
[0131] Specifically, the detailed remaining useful life prediction process is as follows:
[0132] Step 7.1: Input the features and labels of historical data to train the degradation model based on the convolutional gated neural network, and enable the dropout mechanism during the training process;
[0133] Step 7.2: In the prediction, the dropout mechanism is also enabled, and the Monte Carlo simulation method is used to predict the remaining lifetime, which can also obtain the distribution of the remaining lifetime.
[0134] Step 8: Based on the stacking fusion strategy, construct a neural network model that integrates knowledge and data.
[0135] Specifically, such as Figure 3 As shown, the established fusion model is explained in detail below:
[0136] Step 8.1: The network input is a 2D lifetime distribution, including the lifetime distribution predicted based on a statistical model and the lifetime distribution obtained based on a data-driven method. The distribution is discretized and then used as the input.
[0137] Step 8.2: The output is the fusion result of the two, which is a more accurate remaining lifetime prediction result.
[0138] Step 8.3: The hidden layers of the network are alternately connected by CNN layers and fully connected layers. CNN layers can capture short-range spatial relationships, dropout layers are used to prevent overfitting, and fully connected layers are used to fuse features obtained from convolutional layers to obtain the final prediction result.
[0139] Step 9: Integrate the statistical model obtained in Steps 3-5 with the data-driven model obtained in Steps 6 and 7, and use the integrated data as input to train the neural network model of knowledge and data fusion.
[0140] Step 10: For the newly input sample data, update the statistical model online and estimate the lifetime distribution, and use the trained data-driven model to predict the lifetime distribution; use the trained knowledge and data fusion model to fuse the results of the two.
[0141] Step 11: Based on the fusion results in Step 10, calculate the weights and sample the statistical model and the data-driven model respectively to obtain a more accurate lifetime distribution.
[0142] Specifically, such as Figure 4 As shown, the lifetime distribution generation process is as follows:
[0143] Step 11.1: Based on the knowledge and data fusion model, combined with the complementary filter model... Calculate the sampling weights
[0144] Step 11.2: Sample the statistical model and the data-driven model according to the weights respectively. Figure 4 The dashed and solid lines represent two different distributions. By sampling with weights, a more accurate lifetime distribution can be obtained after fusion. Depending on the range of sampling weights, the original distribution can be enhanced or weakened before sampling and superposition.
[0145] This invention discloses a knowledge- and data-fusion method for predicting the remaining lifetime of power systems. It establishes a power system capacity degradation model based on the Wiener process to characterize the degradation state of individual systems; and a power system degradation model based on a gated convolutional neural network to uncover common degradation characteristics of power systems. The knowledge-based statistical model possesses prior knowledge and strong interpretability; the data-based deep learning model mines degradation features through data. Finally, the remaining lifetime distribution is generated using the fusion model and weighted sampling method.
[0146] This invention addresses the shortcomings of existing methods by proposing a knowledge and data fusion-based method for predicting the remaining lifetime of power systems. Based on statistical modeling of the power system, a gated convolutional neural network is used to analyze the system's operating data and predict its remaining lifetime. The two models are then fused using a convolutional neural network, aiming to minimize the error between the predicted distribution mean and the true value. Dynamic weights are sampled from both distributions to complete the prediction of the remaining lifetime distribution of the power system.
[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting the remaining lifetime of a power system by fusing knowledge and data, characterized in that, Includes the following steps: Step S1: Accumulate timing data from the historical operation of the power system and from the cycle life test; Step S2: Preprocess and extract features from the time-series data of the power system during its historical operation to obtain data with degradation trends; Step S3: Establish a power system capacity degradation model based on the Wiener process, which is the statistical model; Step S4: Based on the accumulated data on the power system capacity degradation trend, solve for the parameters in the power system capacity degradation model; Step S5: Calculate the first arrival time and the remaining lifetime distribution of the power system based on the industry power system capacity failure threshold. Step S6: Construct a degradation model based on a convolutional gated recurrent neural network, with historical monitoring data as input and remaining lifetime as output; Step S7: Train a degradation model based on a convolutional gated recurrent neural network to obtain a data-driven model, and estimate the remaining lifetime distribution using the Monte Carlo method; Step S8: Based on the stacking fusion strategy, construct a neural network model that fuses knowledge and data; Step S9: Integrate the relevant results of the statistical model obtained in steps S3-S5 and the data-driven model obtained in steps S6-S7, and use the integrated sample data as input to train the neural network model of knowledge and data fusion. Step S10: For the newly input sample data, update the statistical model online and estimate the remaining lifetime distribution, and use the trained data-driven model to predict the remaining lifetime distribution; use the trained knowledge and data fusion neural network model to fuse the results of the two. Step S11: Based on the fusion results in step S10, calculate the weights and sample the statistical model and the data-driven model respectively to obtain a more accurate remaining lifetime distribution.
2. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 1, characterized in that, Step S2 specifically includes: Preprocessing of time series data includes filling in missing time series data and normalization. The variance statistic is used to determine whether a performance parameter has a trend of change. The variance calculation formula is: Var(X) = E[XE(X)] 2 X is the variable to be judged. Based on the charging and discharging principle of the power system, the changes in the parameter curves are observed, and the performance parameters with degradation trends are selected as the modeling objects.
3. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 2, characterized in that, Step S3 specifically includes: Construct a power system capacity degradation model based on the Wiener process: Y(t)=y0+λt+σ B ·B(t); The power system capacity degradation model is discretized and transformed into the following state equation form: In the formula, λ is the drift coefficient of the degradation process, λ0 follows a normal distribution N(μ0,σ0), where μ0 and σ0 are latent variables to be estimated; ε is the noise of the random walk process, following a normal distribution N(0,Q), where Q is a latent variable to be estimated; y is the degradation amount under study; t is the time of monitoring data; σ B is the coefficient of Brownian motion, is the latent variable to be estimated; B is Brownian motion.
4. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 3, characterized in that, Step S4 specifically includes: The drift coefficients are estimated using the Kalman filter algorithm; The parameters in the power system capacity degradation model are solved iteratively using the EM algorithm, where the maximum likelihood formula is: In the formula, θ=[μ0,σ0,Q,σ B [Y] represents the latent variable to be estimated; 0:i =[y0,y1,…y i ] represents the set of degradation values y at each time step; Λ i =[λ0,λ1,…,λ i ] represents the drift coefficient λ at each time step. i The set of values; finally, the power system capacity degradation model based on the Wiener process is solved.
5. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 4, characterized in that, Step S5 specifically includes the following steps: The power system failure threshold is 30% capacity loss. Based on the definition of first-arrival time and the inverse Gaussian distribution, the lifetime distribution is calculated as follows: In the formula R i Let r be the set of remaining lifetimes. i Let w be a certain estimated value of the remaining lifetime, and w be the failure threshold of the power system capacity. Substitute the relevant variables from the power system capacity degradation model into the equation to obtain the remaining lifetime distribution.
6. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 5, characterized in that, Step S6 specifically includes: The network input consists of voltage, current, and temperature data from historical monitoring data, and the network output is the remaining lifetime. The input time step is m, and the remaining lifetime is predicted using convolutional layers and gated recurrent neural network layers; The hidden layers of the network consist of convolutional neural networks, gated recurrent neural networks, and fully connected layers.
7. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 6, characterized in that, Step S7 specifically includes: Based on the degradation model derived from S6 using a convolutional gated recurrent neural network, the lifetime distribution of the data-driven model is obtained through Monte Carlo dropout simulation.
8. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 7, characterized in that, Step S8 specifically includes: A neural network model for knowledge and data fusion is constructed based on convolutional neural networks and fully connected neural networks; The network's inputs are the outputs of the statistical model and the data-driven model, and the network's output is the predicted mean of the remaining lifetime distribution. The hidden layers of the network include convolutional neural networks and fully connected neural networks.
9. The method for predicting the remaining lifetime of a power system based on knowledge and data fusion according to claim 8, characterized in that, Step 11 specifically includes: Based on the neural network model that integrates knowledge and data, and combined with the complementary filter model, the sampling weights are calculated. By sampling the statistical model and the data-driven model according to the weights, a more accurate distribution of remaining lifetime is obtained.